The effects of biological knowledge graph topology on classical embedding-based link prediction

Area of Science:

• Bioinformatics
• Genomics
• Computational Biology

Background:

• Knowledge graphs (KGs) are crucial for understanding rare diseases and inferring gene-disease links due to limited data.
• Classical knowledge graph embedding methods are effective for biomedical link prediction but underutilized for rare disease variant prioritization.

Purpose of the Study:

• To investigate the impact of knowledge graph topology on the performance of classical knowledge graph embedding for rare disease variant prioritization.
• To challenge the notion that larger, aggregated knowledge graphs are always superior for rare disease research.

Main Methods:

• Utilized the Monarch knowledge graph as a case study.
• Applied classical knowledge graph embedding techniques to predict gene-disease links.
• Evaluated model performance using filtered subsets of the Monarch knowledge graph.

Main Results:

• A filtered Monarch knowledge graph, reduced to 11% of its original size, significantly improved predictive performance.
• Knowledge graph optimization for rare disease variant prioritization is more dependent on data quality than data quantity.

Conclusions:

• Filtering knowledge graphs to retain high-quality information enhances predictive accuracy for rare disease variant prioritization.
• The study underscores the importance of strategic data selection over data aggregation in biomedical knowledge graph applications.